X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?p=dagnn.git;a=blobdiff_plain;f=dagnn.lua;h=b82398c0de429bd19875776fa465222a84504bbe;hp=2ed25c531bd9c02047e33f302a274b18f1b5439a;hb=HEAD;hpb=60568def49e4c624e54f53b4be5783d6cfbe1ea9

diff --git a/dagnn.lua b/dagnn.lua
index 2ed25c5..b82398c 100755
--- a/dagnn.lua
+++ b/dagnn.lua
@@ -1,4 +1,23 @@
 
+--[[
+
+   Copyright (c) 2016 Idiap Research Institute, http://www.idiap.ch/
+   Written by Francois Fleuret <francois.fleuret@idiap.ch>
+
+   This file is free software: you can redistribute it and/or modify
+   it under the terms of the GNU General Public License version 3 as
+   published by the Free Software Foundation.
+
+   It is distributed in the hope that it will be useful, but WITHOUT
+   ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
+   or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public
+   License for more details.
+
+   You should have received a copy of the GNU General Public License
+   along with this file.  If not, see <http://www.gnu.org/licenses/>.
+
+]]--
+
 require 'torch'
 require 'nn'
 
@@ -6,52 +25,140 @@ local DAG, parent = torch.class('nn.DAG', 'nn.Container')
 
 function DAG:__init()
    parent.__init(self)
-   self.pred = {}
-   self.succ = {}
+   -- Nodes are indexed by the module they contain
+   self.node = {}
 end
 
-function DAG:addEdge(a, b)
-   self.sorted = nil
-   local pred, succ = self.pred, self.succ
-   if not pred[a] and not succ[a] then
-      self:add(a)
-   end
-   if not pred[b] and not succ[b] then
-      self:add(b)
-   end
-   pred[b] = pred[b] or {}
-   pred[b][#pred[b] + 1] = a
-   succ[a] = succ[a] or {}
-   succ[a][#succ[a] + 1] = b
-end
-
--- Apply f on t recursively; use the corresponding a1 and a2 elements
--- (i.e. same keys) as second and third parameters to f when
--- available; return the results from f, organized in a similarly
--- nested table.
-function DAG:applyOnModules(f, t, a1, a2)
+-- Apply f on t recursively; use the corresponding elements from args
+-- (i.e. same keys) as second parameter to f when available; return
+-- the results from f, organized in a similarly nested table.
+function DAG:nestedApply(f, t, args)
    if torch.type(t) == 'table' then
       local result = {}
       for k, s in pairs(t) do
-         result[k] = self:applyOnModules(f, s, a1 and a1[k], a2 and a2[k])
+         result[k] = self:nestedApply(f, s, args and args[k])
       end
       return result
    else
-      return f(t, a1, a2)
+      return f(t, args)
+   end
+end
+
+function DAG:createNode(nnm)
+   if not self.node[nnm] then
+      self:add(nnm) -- Add it to the object as a Container
+      local node = {}
+      node.succ = {}
+      node.pred = {}
+      node.index = #self.modules
+      self.node[nnm] = node
+   end
+end
+
+function DAG:putInOrder()
+   if self.sorted then
+      return
+   end
+
+   local distance = {}
+   self:nestedApply(
+      function(m) distance[m] = 1 end,
+      self.inputModules
+   )
+
+   local nc
+   local nl = 0
+   repeat
+      assert(nl < #self.modules, 'Cycle detected in the graph.')
+      nc = 0
+      for nnma, node in pairs(self.node) do
+         for _, nnmb in pairs(node.succ) do
+            if distance[nnma] and (not distance[nnmb] or distance[nnmb] < distance[nnma] + 1) then
+               distance[nnmb] = distance[nnma] + 1
+               nc = nc + 1
+            end
+         end
+      end
+      nl = nl + 1
+   until nc == 0
+
+   for _, nnm in pairs(self.modules) do
+      assert(distance[nnm], 'Some modules are not connected to inputs.')
+   end
+
+   self.sorted = {}
+   for m, d in pairs(distance) do
+      table.insert(self.sorted, { distance = d, nnm = m })
+   end
+
+   table.sort(self.sorted, function(a, b) return a.distance < b.distance end)
+
+   for i, a in ipairs(self.sorted) do self.sorted[i] = a.nnm end
+end
+
+-- This accumulates x in a, where they are both nested tables of
+-- tensors with same structures / keys. If first is true, set a = x
+-- (in which case a can be nil) otherwise a = a + x. The behavior is
+-- undefined if a and x do not have the exact same structure.
+function DAG:nestedAccTensor(a, x, first)
+   if torch.type(x) == 'table' then
+      local b = {}
+      for i in pairs(x) do
+         b[i] = self:nestedAccTensor(a[i], x[i], first)
+      end
+      a = b
+   else
+      if first then
+         if a then
+            a:resizeAs(x):copy(x)
+         else
+            a = x:clone()
+         end
+      else
+         a:add(x)
+      end
    end
+   return a
+end
+
+function DAG:updateGradOutput(node)
+   local gradInputSucc = node.gradInputSucc
+   if #gradInputSucc == 1 then
+      node.gradOutput = gradInputSucc[1]
+   elseif #gradInputSucc > 1 then
+      for k = 1, #gradInputSucc do
+         node.gradOutput = self:nestedAccTensor(node.gradOutput, gradInputSucc[k], k == 1)
+      end
+   end
+end
+
+----------------------------------------------------------------------
+
+-- Connect a sequence of modules
+function DAG:connect(...)
+   self.sorted = nil
+   local prev
+   for _, nnm in pairs({...}) do
+      self:createNode(nnm)
+      if prev then
+         table.insert(self.node[nnm].pred, prev)
+         table.insert(self.node[prev].succ, nnm)
+      end
+      prev = nnm
+   end
+end
+
+function DAG:setLabel(nnm, label)
+   self.node[nnm].label = label
 end
 
 function DAG:setInput(i)
    self.sorted = nil
    self.inputModules = i
-   self:applyOnModules(
-      function(m)
-         if not self.succ[m] or #self.succ[m] == 0 then
-            error('Input modules must have outgoing  edges.')
-         end
-         if self.pred[m] and #self.pred[m] > 0 then
-            error('Input modules cannog have incoming edges.')
-         end
+   self:nestedApply(
+      function(nnm)
+         assert(#self.node[nnm].succ > 0, 'Input modules must have outgoing edges.')
+         assert(#self.node[nnm].pred == 0, 'Input modules cannot have incoming edges.')
       end,
       self.inputModules
    )
@@ -60,110 +167,203 @@ end
 function DAG:setOutput(o)
    self.sorted = nil
    self.outputModules = o
-   self:applyOnModules(
-      function(m)
-         if not self.pred[m] or #self.pred[m] == 0 then
-            error('Output module must have incoming edges.')
-         end
-         if self.succ[m] and #self.succ[m] > 0 then
-            error('Output module cannot have outgoing edges.')
-         end
+   self:nestedApply(
+      function(nnm)
+         assert(#self.node[nnm].pred > 0, 'Output module must have incoming edges.')
+         assert(#self.node[nnm].succ == 0, 'Output module cannot have outgoing edges.')
       end,
       self.outputModules
    )
 end
 
-function DAG:sort()
-   if self.sorted then
-      return
-   end
+function DAG:print()
+   self:putInOrder()
 
-   local distance = {}
+   for i, d in ipairs(self.sorted) do
+      local decoration = ''
+      if self.node[d].label then
+         decoration = ' [' .. self.node[d].label .. ']'
+      end
+      print('#' .. i .. ' -> ' .. torch.type(d) .. decoration)
+   end
+end
 
-   self:applyOnModules(function(m) distance[m] = 1 end, self.inputModules)
+----------------------------------------------------------------------
 
-   local nc
+function DAG:saveDot(filename)
+   local file = (filename and io.open(filename, 'w')) or io.stdout
 
-   repeat
-      nc = 0
-      for i, isucc in pairs(self.succ) do
-         for _, j in pairs(isucc) do
-            if distance[i] and (not distance[j] or distance[j] < distance[i] + 1) then
-               distance[j] = distance[i] + 1
-               nc = nc + 1
-            end
+   local function writeNestedCluster(prefix, list, indent)
+      local indent = indent or ''
+      if torch.type(list) == 'table' then
+         file:write(indent .. '  subgraph cluster_' .. prefix .. ' {\n');
+         for k, x in pairs(list) do
+            writeNestedCluster(prefix .. '_' .. k, x, '  ' .. indent)
          end
+         file:write(indent .. '  }\n');
+      else
+         file:write(indent .. '  ' .. self.node[list].index .. ' [color=red]\n')
       end
-   until nc == 0
-
-   self.sorted = { }
-   for i, d in pairs(distance) do
-      table.insert(self.sorted, { d, i })
    end
 
-   table.sort(self.sorted, function(a, b) return a[1] < b[1] end)
-   for i, a in ipairs(self.sorted) do self.sorted[i] = a[2] end
-end
+   file:write('digraph {\n')
 
-function DAG:print()
-   self:sort()
+   file:write('\n')
 
-   for i, d in ipairs(self.sorted) do
-      print('#' .. i .. ' -> ' .. torch.type(d))
+   writeNestedCluster('input', self.inputModules)
+   writeNestedCluster('output', self.outputModules)
+
+   file:write('\n')
+
+   for nnmb, node in pairs(self.node) do
+      file:write(
+         '  '
+            .. node.index
+            .. ' [shape=box,label=\"' .. (self.node[nnmb].label or torch.type(nnmb)) .. '\"]'
+            .. '\n'
+      )
+
+      for i, nnma in pairs(node.pred) do
+         local decoration = ''
+         if #node.pred > 1 then
+            -- decoration = ' [headlabel=\"' .. i .. '\"]'
+            decoration = ' [label=\"' .. i .. '\"]'
+         end
+         file:write(
+            '  '
+               .. self.node[nnma].index
+               .. ' -> '
+               .. self.node[nnmb].index
+               .. decoration
+               .. '\n'
+         )
+      end
+
+      file:write('\n')
    end
+
+   file:write('}\n')
+
 end
 
+----------------------------------------------------------------------
+
 function DAG:updateOutput(input)
-   self:sort()
-
-   self:applyOnModules(function(m, i) m:updateOutput(i) end, self.inputModules, input)
-
-   for _, d in ipairs(self.sorted) do
-      if self.pred[d] then
-         if #self.pred[d] == 1 then
-            d:updateOutput(self.pred[d][1].output)
-         elseif #self.pred[d] > 1 then
-            local c = {}
-            for k = 1, #self.pred[d] do
-               c[k] = self.pred[d][k].output
+   self:putInOrder()
+
+   self:nestedApply(
+      function(nnm, i)
+         local node = self.node[nnm]
+         node.input = i
+         self:rethrowErrors(nnm, node.index, 'updateOutput', i)
+      end,
+      self.inputModules,
+      input
+   )
+
+   for _, nnm in ipairs(self.sorted) do
+      local node = self.node[nnm]
+      local pred = node.pred
+      if #pred > 0 then
+         local i
+         if #pred == 1 then
+            i = pred[1].output
+         elseif #pred > 1 then
+            i = {}
+            for k = 1, #pred do
+               i[k] = pred[k].output
             end
-            d:updateOutput(c)
          end
+         node.input = i
+         self:rethrowErrors(nnm, node.index, 'updateOutput', i)
       end
    end
 
-   self.output = self:applyOnModules(function(m) return m.output end, self.outputModules)
+   self.output = self:nestedApply(
+      function(m) return m.output end,
+      self.outputModules
+   )
 
    return self.output
 end
 
 function DAG:updateGradInput(input, gradOutput)
-   self:sort()
-
-   self:applyOnModules(function(m, i, go) m:updateGradInput(i, go) end, self.outputModules, input, gradOutput)
-
-   for k = self.sorted, 1, -1 do
-      local m = sorted[k]
-      if self.succ[d] then
-         if #self.succ[d] == 1 then
-            d:updateGradInput(self.succ[d][1].gradInput)
-         elseif #self.succ[d] > 1 then
-            local sum
-            for k = 1, #self.succ[d] do
-               if sum then
-                  sum:add(self.succ[d][k].gradInput)
-               else
-                  sum = self.succ[d][k].gradInput:clone()
-               end
-            end
-            d:updateGradInput(sum)
+   assert(self.sorted, 'There has been a structure change before a DAG:updateGradInput.')
+
+   self:nestedApply(
+      function(nnm, go)
+         local node = self.node[nnm]
+         node.gradOutput = go
+         self:rethrowErrors(nnm, node.index, 'updateGradInput', node.input, go)
+      end,
+      self.outputModules, gradOutput
+   )
+
+   self:nestedApply(
+      function(nnm, i) self.node[nnm].input = i end,
+      self.inputModules, input
+   )
+
+   for _, node in pairs(self.node) do
+      node.gradInputSucc = {}
+   end
+
+   for k = #self.sorted, 1, -1 do
+      local nnm = self.sorted[k]
+      local node = self.node[nnm]
+      local pred = node.pred
+
+      if #node.gradInputSucc > 0 then
+         self:updateGradOutput(node)
+         self:rethrowErrors(nnm, node.index, 'updateGradInput', node.input, node.gradOutput)
+      end
+
+      -- We fill the gradInputSucc of our predecessors
+      if #pred == 1 then
+         table.insert(self.node[pred[1]].gradInputSucc, nnm.gradInput)
+      elseif #pred > 1 then
+         assert(torch.type(nnm.gradInput) == 'table',
+                'Should have a table gradInput since it has multiple predecessors.')
+         for n = 1, #pred do
+            table.insert(self.node[pred[n]].gradInputSucc, nnm.gradInput[n])
          end
       end
    end
 
-   self.gradInput = self:applyOnModules(function(m) return m.gradInput end, self.inputModules)
+   self.gradInput = self:nestedApply(
+      function(m) return m.gradInput end,
+      self.inputModules
+   )
 
    return self.gradInput
 end
 
-return DAG
+function DAG:accGradParameters(input, gradOutput, scale)
+   assert(self.sorted, 'There has been a structure change before a DAG:accGradParameters.')
+
+   self:nestedApply(
+      function(nnm, go) self.node[nnm].gradOutput = go end,
+      self.outputModules, gradOutput
+   )
+
+   self:nestedApply(
+      function(nnm, i) self.node[nnm].input = i end,
+      self.inputModules, input
+   )
+
+   for k = 1, #self.modules do
+      local nnm = self.modules[k]
+      local node = self.node[nnm]
+      self:rethrowErrors(nnm, k, 'accGradParameters', node.input, node.gradOutput, scale)
+   end
+end
+
+function DAG:clearState()
+   self.sorted = nil
+   for _, node in pairs(self.node) do
+      node.input = nil
+      node.gradInputSucc = nil
+      node.gradOutput = nil
+   end
+   return parent.clearState(self)
+end